Ten Yrs from now, when only AI codes, what's the stack?

At that point you wouldn't care what the ingredients were.

I guess we go back to basics since the main reason to use frameworks and stacks are the benefits of like, consistent structure, knowledge shared across teams, hiring people with prior experience from other jobs. Those don't matter at all to AI.

What does matter is poisoned dependencies, it might end up better to have AI generating bespoke routing, data/state stores, ORMs etc that would be redundant for you to get distracted with today. Humans can barely police dependencies today when humans are producing the dependencies and curating their use so probably a 10x bigger problem in a decade.

Why have a stack, machine-code all the way down baby

10 years from now? It's possible AI (whatever that means then) is essentially responsible for management of a systems state. At that point it's not so much code and more like an oracle that understands what questions/mutations have been asked of it based on how it's been instructed to manage the system.

But if we are still dealing with code... not sure. I think languages could stand to benefit from memory of a sorts, which is to say code executes in strange ways (still probably with "AI in the loop") and execution paths are aware of the history of code that preceded it and essentially picks the optimal path -- if new code executes and doesn't satisfy some conditions, it falls back to old code. Again, driven by real time "AI" evaluation or some predetermined expected outcomes. In this way progressive delivery is a function of the language in concert with the AI that wrote the code, not some external rollout and monitoring system.